Text Region Adaptive Multi-temporal Stone Inscriptions Image Registration Framework

Abstract. Multi-temporal comparative analysis of stone inscription images helps understand weathering processes and enables more targeted conservation measures, with accurate image registration serving as the prerequisite for such analysis. However, existing registration methods perform poorly on these images due to sparse and unevenly distributed feature points in text regions caused by natural weathering processes such as weathering and exfoliation. This study proposes a semantic-aware registration framework for multi-temporal stone inscription images, implementing a four-stage processing workflow: global initial registration using Scale-Invariant Feature Transform (SIFT) algorithm with geometric filtering; region partitioning through grid division and text region identification; sub-block enhancement applying SIFT-Accelerated KAZE (AKAZE) dual-feature fusion strategy for regions with insufficient matching points; and global fine registration integrating multi-source matching points with Random Sample Consensus (RANSAC) optimization. Experiments on stone inscriptions with varying degrees of weathering validate the effectiveness of this method: for well-preserved samples, the Mean Euclidean Error (MEE) decreased by 94.8% and structural similarity improved by 813.7%; for severely weathered samples, matching points in text regions increased 18-fold with an inlier ratio reaching 97.35%. This research represents the integration of text region semantic recognition with adaptive feature enhancement, providing reliable technical support for multi-temporal stone inscription analysis and cultural heritage conservation.